ACADEMIA

Researchers Simulate Individual Thunderstorms in Real Time, Arming Forecasters With Better Guidance for Storm Watches and Warnings: As America's autumn storm season for 2006 continues, the nation's weather forecasters have a relentless demand for reliable information that can help them warn citizens of potentially devastating storms. Providing early warning of tornadoes, thunderstorms, blizzards and hail storms can save lives and protect property. No one knows this better than researchers at the National Severe Storms Laboratory (NSSL) in Norman, Okla., who work side-by-side with National Weather Service (NWS) forecasters every day. NSSL recently acquired a powerful new SGI Altix computer from SGI to generate exceptionally detailed computer simulations of U.S. weather systems. On Altix, NSSL is producing experimental daily forecasts with a "next generation" configuration of the new Weather Research and Forecasting (WRF) model -- a configuration that provides three times the spatial resolution of forecasts that are disseminated daily by current operational NWS models. NSSL has configured the WRF model to blanket most of the continental U.S. with data points every 4 kilometers (about 2.5 miles). In addition, they provide 35 vertical layers of detail. These high-resolution forecasts are generated and evaluated as part of a collaborative effort involving NSSL, the NWS, NASA, and the University of Alabama at Huntsville. NSSL programmed the Altix system to crank out a high-resolution 36-hour forecast every day, and it has proven to be exceptionally reliable at this task. "The Altix system allows us to reach the cutting edge in terms of high- resolution forecasts over a large domain," said Jack Kain, research meteorologist at NSSL, part of the National Oceanic and Atmospheric Administration (NOAA). "Our goal is to prepare forecasters for the next generation of operational forecast models while providing forecaster feedback to model developers." SGI Altix Delivers for NSSL NSSL's cutting-edge WRF research is made possible by a 64-processor SGI Altix 3700 Bx2 system with 128GB of memory and running SUSE® Linux Enterprise Server 10 from Novell. While Altix handles both distributed- and shared-memory applications, NSSL runs WRF in MPI mode, assigning specific tasks to each of the system's 64 Intel Itanium 2 processors to optimize production of the time-sensitive weather forecasts. With these runtime parameters, the Altix produces the daily high resolution 36-hour forecasts in about seven to eight hours, providing timely delivery to forecasters and leaving computer time to support related research and model development on the same system. NSSL acquired the SGI Altix system specifically for the project because the new WRF model required far more computing resources than NSSL's existing systems could deliver. "We have not had the capability to produce real-time operational forecasts like this for four or five years," said Kain. "We had a 40-processor cluster, but it simply was not as reliable or as fast as this Altix." "Every time you double the resolution of this model, you increase the computational requirements by at least a factor of ten," explained Louis Wicker, research scientist at NSSL. "Running the WRF model at such high resolution requires one to represent the complex interaction between winds, pressure, temperature, ice, and liquid water that exists in and around convective clouds, and you need a computer system that can handle the computational demands of those details." The ability of NSSL's Altix system to operate even distributed memory applications like WRF under a single system image (SSI) of Linux has proven advantageous as well. "A typical cluster might have 64 nodes with 64 copies of Linux running on them," said Wicker. "And yes, the WRF model will run on such a system reasonably well. But we gain much more with the Altix, because the SSI reduces the complexity of the system while maintaining the performance. "For example, a cluster solution would require the use of a parallel file system to meet the level of I/O performance needed for these forecasts. But with the Altix system, we have sufficient I/O performance for WRF. The amount of system administration associated with maintaining all those copies of Linux is reduced from 64 to one. With SGI, we not only gain performance, but we gain simplicity as well. That actually lowers our total cost of ownership." Delivering performance has been a key focal point since NSSL first began the search for its latest system. Rather than promote a specific configuration to NSSL, said Kain, SGI focused instead on delivering the level of performance NSSL needed. "We didn't have a guaranteed level of performance from the vendor who sold us the cluster we were using before, but with SGI, we do. That has made a big difference. We've been very happy with Altix and the SGI support that is part of the package." "We set out from the beginning to find a solution that would allow our scientists to concentrate on accomplishing the science, without having to worry about the details of the computer infrastructure," said Kevin Kelleher, NSSL Chief Information Officer. "So far, SGI has delivered such a solution." Installed in May and brought fully online this summer after NSSL moved into new facilities, the new Altix system is also used for other NSSL research, including a separate effort to produce short-term forecasts of severe thunderstorms and tornadoes. Part of a NOAA/National Weather Service initiative called Warn on Forecast, the project is aimed at using ultra-high- resolution models (perhaps eight times more detailed than today's 4 km WRF model) to forecast severe weather events an hour or more before they hit a specific location. The project is expected to last up to a decade. Fortunately for NSSL, SGI Altix is built on a massively scalable architecture, and today supports up to 1,024 processors under one Linux SSI. The WRF project at NSSL is also unusual because it is the product of ongoing collaboration between researchers and practitioners (operational forecasters). "It's challenging to collaborate with operational forecasters, because they have their own priorities and usually don't have much time to do research," said Kain. "But the situation in Norman (OK) is unique. At the new National Weather Center we have a thriving collaborative environment that promotes interactions between model developers and forecasters who use the models. I'm convinced this leads to more efficient model development and a better end product. Ultimately, this means better weather forecasts for the public."